VAE Latent Traversal for Radar Clutter Suppression

Radar systems operating in complex modern environments often encounter significant signal back-scatter from the physical scene, resulting in unwanted clutter returns. Depending on ambient characteristics of these environments, this clutter can be strong and complex, making it challenging to model and suppress without simultaneously impacting targets. In this paper, we propose an unsupervised deep learning method for radar clutter suppression without the need for paired data. Using a variational autoencoder (VAE), we learn to reconstruct unpaired sets of cluttered and clean range-Doppler maps, then solve and apply the mean latent traversal vector between them to suppress clutter. Experiments on real-world data from two urban sites showed that our method improves the detection performance of a known target compared to the unsuppressed baseline. We also explore transferring the learnt network and vector from one site to another site, and conclude that the method still offers improvements over the unsuppressed baseline, but yields best results when trained and applied on the same site. Further analysis on mismatched clutter and target strengths using simulated sea clutter also showed that best results were obtained when both target and clutter strengths were high.